Fix case of nested tensors in BatchMixFeature

[Rocketknight1]

Huge kudos to @zucchini-nlp for figuring this one out - we were getting intermittent failures and I couldn't figure out why. The reason was that BatchMixFeature.to() was breaking the structure of some inputs, because it used a nested list comprehension that assumed a specific list depth that wasn't always respected by PixtralProcessor.

This PR replaces the slightly hacky .to() code with a recursion that should preserve nesting structure regardless of list depth.

TODO:

• Compare against the VLLM implementation of Pixtral to see how they handle this (if their code supports batched?)

[qubvel]

Hey @Rocketknight1! There is a PR that might be relevant in terms of how args/kwargs are propagated

• Option to set 'non_blocking' for to(device) in BatchEncoding and BatchFeature #34883

P.S. didn't notice it's a model-related change only, so might be not relevant

[Rocketknight1]

@qubvel thanks for letting me know regardless, since we might make the same change in BatchFeature! The two PRs shouldn't conflict much, though - if you merge yours first, I can just rebase and incorporate those changes afterwards.

[Rocketknight1]

Note - this PR should not be merged yet, as I believe the issue in .to() was actually masking the fact that Processor outputs don't match the expectations of our Pixtral modelling code!

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[Rocketknight1]

cc @ArthurZucker @LysandreJik for core maintainer review, this one should be ready!

[ArthurZucker]

Thanks for fixing!

[ArthurZucker]

these tests should be in the image processing / processing but alrights, thanks! 🤗

[ArthurZucker]

it's actually cleaner!

[Rocketknight1]

Closing because this change is included in #34801 instead